pumaDB pumaDB vs Zep

pumaDB vs Zep

Explicit JSON memory, or governed context graphs?

pumaDB and Zep both help agents keep useful context, but they sit at very different layers. pumaDB is a hosted JSON memory API and MCP server. Zep is enterprise agent memory infrastructure built around temporal context graphs, assembled context, provenance, and governance.

pumaDB

Hosted JSON memory API

vs
Zep

Context graph memory

Quick read

Choose pumaDB when an agent or app needs simple explicit state it can inspect, update, and restore. Choose Zep when a production agent needs graph-built context from chat history, business data, user behavior, and changing facts.

Zep is context infrastructure. pumaDB is the small memory primitive.

Best fit

Use the tool that matches the size of the job.

Choose pumaDB for

  • Hosted MCP memory without running a database, vector index, or graph stack
  • Small structured JSON tables with explicit app-defined records
  • Server-side REST calls from Workers, API routes, CLIs, and scripts
  • Reviewable memory with row versions, restore, and explicit cleanup
  • Typed safe memory for resources, code, Markdown, commands, and config

Choose Zep for

  • Enterprise agents that need temporal knowledge graphs
  • Context assembled from chat history, business data, user interactions, JSON, documents, and events
  • Graph RAG and retrieval over entities, facts, relationships, and historical changes
  • Provenance, access control, retention, and audit requirements around agent context
  • Managed context graph infrastructure, Graphiti, SDKs, and enterprise deployment options

Detailed comparison

Where they differ.

Area pumaDB Zep
Primary job A small hosted JSON database for agent memory and lightweight server-side app state. Enterprise agent memory infrastructure for building, governing, and retrieving context from temporal graphs.
Data model Named JSON tables. Current account limits are 20 tables, 1,000 rows per table, 25 MB per account, and 64 KB per row. Temporal context graphs made from entities, facts, relationships, episodes, validity windows, and source provenance.
Write behavior Agents and apps explicitly add, upsert, patch, batch, or delete JSON rows. Applications ingest messages, text, JSON, business data, documents, and events so Zep can construct and update graph context.
Query behavior CRUD, simple equality filters, sorting, counting, batch writes, and short-lived viewer links for larger results. Retrieves assembled context blocks and graph results based on relevant facts, entities, relationships, and temporal state.
AI layer Stores memory as explicit JSON records. It does not extract facts, build graphs, run Graph RAG, or assemble LLM context. Positions itself around fact extraction, temporal graph updates, context engineering, Graph RAG, and token-efficient context assembly.
Agent interface Hosted Streamable HTTP MCP, local stdio MCP, REST API, and CLI are built into the product surface. SDKs, APIs, examples, and graph or memory APIs for adding data and retrieving context for LLM calls.
Governance Uses account-scoped auth, OAuth for hosted MCP, API keys, explicit delete filters, and archived row restore. Emphasizes enterprise controls such as provenance, access control, retention, audit, trust, and governed context retrieval.
Operational shape Hosted on Cloudflare with deliberately small limits and no graph, vector, or model configuration. A managed enterprise memory platform with context graph infrastructure, Graphiti, and deployment or compliance options.
Recovery model Every update and delete archives the previous row. The last 10 versions are kept for 30 days and can be restored. Tracks changing facts over time with validity and provenance, rather than exposing row-level JSON version restore as the central primitive.

Positioning

A practical way to think about the split.

pumaDB is the memory primitive

It is useful when an agent needs durable state that stays explicit: preferences, task state, project facts, resource references, code snippets, commands, config examples, and small records your own app already understands.

Zep is the context graph system

It is useful when the job is ingesting conversations and business data, extracting changing facts, preserving provenance, applying governance, and retrieving a ready-to-use context block for an agent.

Sources

What this comparison is based on.

Product pages change over time. This page compares pumaDB's current documented behavior in this codebase with Zep's public project documentation.

Zep homeZep docsZep graph overviewZep memory APIpumaDB docspumaDB install

Start small

Add hosted memory without standing up a database.

Connect pumaDB over hosted MCP, or call the REST API from trusted server-side code.

Install pumaDB Read the API docs